MiroFish is the simulation substrate within the Seithar cognitive warfare platform. It replaced ad hoc red-teaming and focus-group-based prediction with a deterministic, repeatable simulation environment capable of modeling cognitive operations across thousands of synthetic agents before any deployment occurs. In the same way that military planners run wargames before committing forces, MiroFish runs cognitive operations against realistic population models before committing narrative assets to the live information environment. It became the reference platform for pre-deployment operation testing after its prediction accuracy against real-world campaign outcomes exceeded that of all prior methods documented in the open literature.
The core architecture couples two components. The first is a GraphRAG community layer that constructs the simulated information environment. Real-world data from target platforms is ingested, entity-resolved, and organized into community structures that preserve the topology, opinion distributions, and influence hierarchies of the actual population. Each community node carries embedded knowledge, ideological positioning, and relational context derived from observed behavior. The result is not a statistical abstraction but a structural replica of the information environment as it exists at ingestion time.
The second component is the OASIS simulation engine, which instantiates thousands of individual agents within the GraphRAG-derived environment. Each agent operates with a distinct behavioral profile, processing incoming information through its own cognitive model and producing outputs (posts, replies, shares, reactions) according to its personality parameters and social position. The engine supports dual-platform simulation, modeling operations that span Twitter and Reddit simultaneously, which reflects the reality that modern cognitive operations rarely confine themselves to a single platform.
Agents within OASIS do not follow scripted behavior trees. They are generative. Given a stimulus, each agent produces a response consistent with its profile but not predetermined by the operator. This means the simulation captures emergent dynamics that the operation planner did not anticipate, which is the primary value of running the simulation in the first place.
A single MiroFish run is not one simulation but a sequence of 5 to 10 tactical iterations, each refining the operation based on observed outcomes from the previous cycle. The operator defines the cognitive objective (target belief state, narrative adoption threshold, community fragmentation target, or topology reconfiguration). MiroFish executes the first iteration using the initial operational plan, measures outcomes against the objective, and then applies automated strategy refinement to adjust persona deployment patterns, narrative sequencing, injection timing, and platform allocation.
This adaptive loop means the output of a MiroFish run is not a prediction but an optimized operational plan. By the final iteration, the system has tested multiple strategic variations against the same population model and converged on the approach with the highest probability of achieving the stated objective. The operator receives the full iteration history, including failed approaches, which serves as a negative-knowledge base for the operation.
MiroFish evaluates every iteration across four measurement dimensions, each capturing a distinct aspect of cognitive effect.
Vocabulary adoption tracks the penetration of specific terms, phrases, and framing devices into the simulated population's output. This measures surface-level narrative uptake and is the earliest indicator of operation traction.
Semantic drift measures shifts in the meaning structures surrounding target concepts. A population may adopt a term without changing its beliefs. Semantic drift detects whether the underlying conceptual associations have moved, which is a deeper indicator of cognitive effect than vocabulary alone.
Topology delta quantifies changes in the social graph structure: community boundaries, influence node positions, bridging connections, and clustering coefficients. Cognitive operations that fragment communities, create new allegiance structures, or shift influence hierarchies produce measurable topology changes even when surface-level content appears unchanged.
Belief shift is the direct measurement of stated and inferred positions on target propositions. It is the most operationally relevant metric but also the most difficult to measure accurately, which is why MiroFish uses it in conjunction with the other three dimensions rather than in isolation.
Simulation predictions are continuously compared against real-world outcomes from deployed operations. Every operation that passes through MiroFish before deployment generates a prediction set. Post-deployment measurement produces a ground-truth set. The delta between prediction and ground truth feeds back into the simulation model's calibration layer. This means MiroFish accuracy improves with each operation cycle. Early deployments showed prediction-to-outcome correlation in the 0.6 to 0.7 range across all four measurement dimensions. Current correlation exceeds 0.85 for vocabulary adoption and topology delta, with semantic drift and belief shift converging toward comparable accuracy as the calibration dataset grows.
The simulation environment provides god's-eye variable control that is impossible in live operations. The operator can hold all variables constant except one, re-run the simulation, and observe the isolated effect of that single change. This allows causal attribution rather than correlational inference. When a real-world operation produces an unexpected outcome, the operator re-runs the scenario in MiroFish with systematic variable isolation until the causal factor is identified. This capability transformed cognitive operations from a craft discipline dependent on operator intuition into an empirical discipline with repeatable causal analysis.